mitai-jinkendo/backend/prompt_executor.py

"""
Unified Prompt Executor (Issue #28 Phase 2)

Executes both base and pipeline-type prompts with:
- Dynamic placeholder resolution
- JSON output validation
- Multi-stage parallel execution
- Reference and inline prompt support
"""
import json
import re
from typing import Dict, Any, Optional
from db import get_db, get_cursor, r2d
from fastapi import HTTPException
from placeholder_resolver import get_catalog_row_for_key


def resolve_placeholders(template: str, variables: Dict[str, Any], debug_info: Optional[Dict] = None, catalog: Optional[Dict] = None) -> str:
    """
    Replace {{placeholder}} with values from variables dict.

    Modifiers (Katalog aus get_placeholder_catalog empfohlen):
    - {{key|d}} — Wert — description (kurz)
    - {{key|x}} — nur Erklärung (Katalogfeld ai_caption), ohne Zahlenwert
    - {{key|d,x}} — Wert — description — Erklärung

    Args:
        template: String with {{key}} or {{key|modifiers}} placeholders
        variables: Dict of key -> value mappings
        debug_info: Optional dict to collect debug information
        catalog: Optional placeholder catalog for descriptions (from get_placeholder_catalog)

    Returns:
        Template with placeholders replaced
    """
    resolved = {}
    unresolved = []

    def replacer(match):
        full_placeholder = match.group(1).strip()

        # Parse key and modifiers (e.g., "weight_aktuell|d" -> key="weight_aktuell", modifiers="d")
        parts = full_placeholder.split('|')
        key = parts[0].strip()
        modifiers = parts[1].strip() if len(parts) > 1 else ''
        mods = {x.strip().lower() for x in modifiers.split(",") if x.strip()}
        want_d = "d" in mods
        want_x = "x" in mods

        def _warn(msg: str):
            if debug_info is not None:
                debug_info.setdefault("warnings", []).append(msg)

        row = get_catalog_row_for_key(catalog, key) if catalog else None

        if want_x and not want_d:
            if key not in variables:
                if debug_info is not None:
                    unresolved.append(key)
                return match.group(0)
            expl = (row.get("ai_caption") or "").strip() if row else ""
            if not expl and catalog is None:
                _warn(f"Modifier |x für {key}: Katalog fehlt (ai_caption).")
            out = expl
            if debug_info is not None:
                resolved[key] = out[:100] + ("..." if len(out) > 100 else "")
            return out

        if key not in variables:
            if debug_info is not None:
                unresolved.append(key)
            return match.group(0)

        value = variables[key]
        if isinstance(value, (dict, list)):
            resolved_value = json.dumps(value, ensure_ascii=False)
        else:
            resolved_value = str(value)

        if not want_d and not want_x:
            out = resolved_value
            if debug_info is not None:
                resolved[key] = out[:100] + ("..." if len(out) > 100 else "")
            return out

        parts = [resolved_value]
        if want_d:
            if row:
                desc = (row.get("description") or "").strip()
                if desc:
                    parts.append(desc)
            else:
                _warn(f"Modifier |d für {key}: Katalog fehlt (description).")
        if want_x:
            expl = (row.get("ai_caption") or "").strip() if row else ""
            if expl:
                parts.append(expl)
            elif catalog is not None:
                _warn(f"Modifier |x (mit |d) für {key}: ai_caption leer.")

        out = " — ".join(parts)
        if debug_info is not None:
            resolved[key] = out[:100] + ("..." if len(out) > 100 else "")
        return out

    result = re.sub(r'\{\{([^}]+)\}\}', replacer, template)

    # Store debug info
    if debug_info is not None:
        debug_info['resolved_placeholders'] = resolved
        debug_info['unresolved_placeholders'] = unresolved

    return result


def validate_json_output(output: str, schema: Optional[Dict] = None, debug_info: Optional[Dict] = None) -> Dict:
    """
    Validate that output is valid JSON.

    Unwraps Markdown-wrapped JSON (```json ... ```) if present.

    Args:
        output: String to validate
        schema: Optional JSON schema to validate against (TODO: jsonschema library)
        debug_info: Optional dict to attach to error for debugging

    Returns:
        Parsed JSON dict

    Raises:
        HTTPException: If output is not valid JSON (with debug info attached)
    """
    # Try to unwrap Markdown code blocks (common AI pattern)
    unwrapped = output.strip()
    if unwrapped.startswith('```json'):
        # Extract content between ```json and ```
        lines = unwrapped.split('\n')
        if len(lines) > 2 and lines[-1].strip() == '```':
            unwrapped = '\n'.join(lines[1:-1])
    elif unwrapped.startswith('```'):
        # Generic code block
        lines = unwrapped.split('\n')
        if len(lines) > 2 and lines[-1].strip() == '```':
            unwrapped = '\n'.join(lines[1:-1])

    try:
        parsed = json.loads(unwrapped)
        # TODO: Add jsonschema validation if schema provided
        return parsed
    except json.JSONDecodeError as e:
        error_detail = {
            "error": f"AI returned invalid JSON: {str(e)}",
            "raw_output": output[:500] + ('...' if len(output) > 500 else ''),
            "unwrapped": unwrapped[:500] if unwrapped != output else None,
            "output_length": len(output)
        }
        if debug_info:
            error_detail["debug"] = debug_info

        raise HTTPException(
            status_code=500,
            detail=error_detail
        )


async def execute_prompt(
    prompt_slug: str,
    variables: Dict[str, Any],
    openrouter_call_func,
    enable_debug: bool = False
) -> Dict[str, Any]:
    """
    Execute a single prompt (base or pipeline type).

    Args:
        prompt_slug: Slug of prompt to execute
        variables: Dict of variables for placeholder replacement
        openrouter_call_func: Async function(prompt_text) -> response_text
        enable_debug: If True, include debug information in response

    Returns:
        Dict with execution results:
        {
            "type": "base" | "pipeline",
            "slug": "...",
            "output": "..." | {...},  # String or parsed JSON
            "stages": [...]  # Only for pipeline type
            "debug": {...}  # Only if enable_debug=True
        }
    """
    # Load prompt from database
    with get_db() as conn:
        cur = get_cursor(conn)
        cur.execute(
            """SELECT * FROM ai_prompts
               WHERE slug = %s AND active = true""",
            (prompt_slug,)
        )
        row = cur.fetchone()
        if not row:
            raise HTTPException(404, f"Prompt nicht gefunden: {prompt_slug}")

        prompt = r2d(row)

    prompt_type = prompt.get('type', 'pipeline')

    # Get catalog from variables if available (passed from execute_prompt_with_data)
    catalog = variables.pop('_catalog', None) if '_catalog' in variables else None

    if prompt_type == 'base':
        # Base prompt: single execution with template
        return await execute_base_prompt(prompt, variables, openrouter_call_func, enable_debug, catalog)

    elif prompt_type == 'pipeline':
        # Pipeline prompt: multi-stage execution
        return await execute_pipeline_prompt(prompt, variables, openrouter_call_func, enable_debug, catalog)

    elif prompt_type == 'workflow':
        # Workflow prompt: graph-based execution (Phase 0: Foundation)
        return await execute_workflow_prompt(prompt, variables, openrouter_call_func, enable_debug, catalog)

    else:
        raise HTTPException(400, f"Unknown prompt type: {prompt_type}")


async def execute_base_prompt(
    prompt: Dict,
    variables: Dict[str, Any],
    openrouter_call_func,
    enable_debug: bool = False,
    catalog: Optional[Dict] = None,
    node_questions: Optional[list] = None  # Phase 1: Knotengebundene Fragen
) -> Dict[str, Any]:
    """
    Execute a base-type prompt (single template).

    Phase 1: Unterstützt Fragenergänzungen (Hybridmodell)
    - node_questions: Knotengebundene Fragen (Priorität 1)
    - prompt.question_augmentations: Prompt-Defaults (Priorität 2)
    """
    from question_augmenter import (
        parse_question_augmentations_from_jsonb,
        merge_question_augmentations,
        augment_prompt_with_questions
    )
    from result_container_parser import parse_result_container_robust

    template = prompt.get('template')
    if not template:
        raise HTTPException(400, f"Base prompt missing template: {prompt['slug']}")

    debug_info = {} if enable_debug else None

    # Phase 1: Load question augmentations (Hybridmodell)
    prompt_default_questions = None
    if prompt.get('question_augmentations'):
        try:
            from workflow_models import QuestionAugmentation
            prompt_default_questions = parse_question_augmentations_from_jsonb(
                prompt['question_augmentations']
            )
        except Exception as e:
            if enable_debug:
                debug_info['question_augmentations_error'] = str(e)

    # Merge question augmentations (Vorrangregel: Knoten > Prompt)
    questions = merge_question_augmentations(node_questions, prompt_default_questions)

    # Resolve placeholders (with optional catalog for |d modifier)
    prompt_text = resolve_placeholders(template, variables, debug_info, catalog)

    # Phase 1: Augment prompt with questions (if any)
    if questions:
        prompt_text = augment_prompt_with_questions(prompt_text, questions)
        if enable_debug:
            debug_info['question_augmentations_count'] = len(questions)
            debug_info['question_types'] = [q.type for q in questions]

    if enable_debug:
        debug_info['template'] = template
        # Volltext für Test-UI (Admin); sehr große Prompts nur weich begrenzen
        _max = 512 * 1024
        debug_info['final_prompt'] = (
            prompt_text if len(prompt_text) <= _max else prompt_text[:_max] + "\n… [gekürzt, >512KB]"
        )
        debug_info['available_variables'] = list(variables.keys())

    # Call AI
    response = await openrouter_call_func(prompt_text)

    if enable_debug:
        debug_info['ai_response_length'] = len(response)
        debug_info['ai_response_preview'] = response[:200] + ('...' if len(response) > 200 else '')

    # Phase 1: Parse structured result if questions were used
    if questions:
        expected_question_types = [q.type for q in questions]
        container = parse_result_container_robust(response, expected_question_types)

        if enable_debug:
            debug_info['parsing_status'] = container['parsing_status']
            debug_info['parsing_warnings'] = container.get('warnings', [])

        output = container
        output_format = 'structured_container'  # New format type
    else:
        # Legacy behavior: Validate JSON if required
        output_format = prompt.get('output_format', 'text')
        if output_format == 'json':
            output = validate_json_output(response, prompt.get('output_schema'), debug_info if enable_debug else None)
        else:
            output = response

    result = {
        "type": "base",
        "slug": prompt['slug'],
        "output": output,
        "output_format": output_format
    }

    if enable_debug:
        result['debug'] = debug_info

    return result


async def execute_pipeline_prompt(
    prompt: Dict,
    variables: Dict[str, Any],
    openrouter_call_func,
    enable_debug: bool = False,
    catalog: Optional[Dict] = None
) -> Dict[str, Any]:
    """
    Execute a pipeline-type prompt (multi-stage).

    Each stage's results are added to variables for next stage.
    """
    stages = prompt.get('stages')
    if not stages:
        raise HTTPException(400, f"Pipeline prompt missing stages: {prompt['slug']}")

    # Parse stages if stored as JSON string
    if isinstance(stages, str):
        stages = json.loads(stages)

    stage_results = []
    context_vars = variables.copy()
    pipeline_debug = [] if enable_debug else None

    # Execute stages in order
    for stage_def in sorted(stages, key=lambda s: s['stage']):
        stage_num = stage_def['stage']
        stage_prompts = stage_def.get('prompts', [])

        if not stage_prompts:
            continue

        stage_debug = {} if enable_debug else None
        if enable_debug:
            stage_debug['stage'] = stage_num
            stage_debug['available_variables'] = list(context_vars.keys())
            stage_debug['prompts'] = []

        # Execute all prompts in this stage (parallel concept, sequential impl for now)
        stage_outputs = {}

        for prompt_def in stage_prompts:
            source = prompt_def.get('source')
            output_key = prompt_def.get('output_key', f'stage{stage_num}')
            output_format = prompt_def.get('output_format', 'text')

            prompt_debug = {} if enable_debug else None

            if source == 'reference':
                # Reference to another prompt
                ref_slug = prompt_def.get('slug')
                if not ref_slug:
                    raise HTTPException(400, f"Reference prompt missing slug in stage {stage_num}")

                if enable_debug:
                    prompt_debug['source'] = 'reference'
                    prompt_debug['ref_slug'] = ref_slug

                # Load referenced prompt
                result = await execute_prompt(ref_slug, context_vars, openrouter_call_func, enable_debug)
                output = result['output']

                if enable_debug and 'debug' in result:
                    prompt_debug['ref_debug'] = result['debug']

            elif source == 'inline':
                # Inline template
                template = prompt_def.get('template')
                if not template:
                    raise HTTPException(400, f"Inline prompt missing template in stage {stage_num}")

                placeholder_debug = {} if enable_debug else None
                prompt_text = resolve_placeholders(template, context_vars, placeholder_debug, catalog)

                if enable_debug:
                    prompt_debug['source'] = 'inline'
                    prompt_debug['template'] = template
                    _max = 512 * 1024
                    prompt_debug['final_prompt'] = (
                        prompt_text if len(prompt_text) <= _max else prompt_text[:_max] + "\n… [gekürzt, >512KB]"
                    )
                    prompt_debug.update(placeholder_debug)

                response = await openrouter_call_func(prompt_text)

                if enable_debug:
                    prompt_debug['ai_response_length'] = len(response)
                    prompt_debug['ai_response_preview'] = response[:200] + ('...' if len(response) > 200 else '')

                # Validate JSON if required
                if output_format == 'json':
                    output = validate_json_output(response, prompt_def.get('output_schema'), prompt_debug if enable_debug else None)
                else:
                    output = response

            else:
                raise HTTPException(400, f"Unknown prompt source: {source}")

            # Store output with key
            stage_outputs[output_key] = output

            # Add to context for next stage
            context_var_key = f'stage_{stage_num}_{output_key}'
            context_vars[context_var_key] = output

            if enable_debug:
                prompt_debug['output_key'] = output_key
                prompt_debug['context_var_key'] = context_var_key
                stage_debug['prompts'].append(prompt_debug)

        stage_results.append({
            "stage": stage_num,
            "outputs": stage_outputs
        })

        if enable_debug:
            stage_debug['output'] = stage_outputs  # Add outputs to debug info for value table
            pipeline_debug.append(stage_debug)

    # Final output is last stage's first output
    final_output = stage_results[-1]['outputs'] if stage_results else {}

    result = {
        "type": "pipeline",
        "slug": prompt['slug'],
        "stages": stage_results,
        "output": final_output,
        "output_format": prompt.get('output_format', 'text')
    }

    if enable_debug:
        result['debug'] = {
            'initial_variables': list(variables.keys()),
            'stages': pipeline_debug
        }

    return result


async def execute_prompt_with_data(
    prompt_slug: str,
    profile_id: str,
    modules: Optional[Dict[str, bool]] = None,
    timeframes: Optional[Dict[str, int]] = None,
    openrouter_call_func = None,
    enable_debug: bool = False
) -> Dict[str, Any]:
    """
    Execute prompt with data loaded from database.

    Args:
        prompt_slug: Slug of prompt to execute
        profile_id: User profile ID
        modules: Dict of module -> enabled (e.g., {"körper": true})
        timeframes: Dict of module -> days (e.g., {"körper": 30})
        openrouter_call_func: Async function for AI calls
        enable_debug: If True, include debug information in response

    Returns:
        Execution result dict
    """
    from datetime import datetime, timedelta
    from placeholder_resolver import get_placeholder_example_values, get_placeholder_catalog

    # Build variables from data modules
    variables = {
        'profile_id': profile_id,
        'today': datetime.now().strftime('%Y-%m-%d')
    }

    # Load placeholder catalog for |d / |x Modifier
    try:
        catalog = get_placeholder_catalog(profile_id)
    except Exception as e:
        catalog = None
        print(f"Warning: Could not load placeholder catalog: {e}")

    variables['_catalog'] = catalog  # Will be popped in execute_prompt (can be None)

    # Add PROCESSED placeholders (name, weight_trend, caliper_summary, etc.)
    # This makes old-style prompts work with the new executor
    try:
        processed_placeholders = get_placeholder_example_values(profile_id)
        # Remove {{ }} from keys (placeholder_resolver returns them with wrappers)
        cleaned_placeholders = {
            key.replace('{{', '').replace('}}', ''): value
            for key, value in processed_placeholders.items()
        }
        variables.update(cleaned_placeholders)
    except Exception as e:
        # Continue even if placeholder resolution fails
        if enable_debug:
            variables['_placeholder_error'] = str(e)

    # Load data for enabled modules
    if modules:
        with get_db() as conn:
            cur = get_cursor(conn)

            # Weight data
            if modules.get('körper'):
                days = timeframes.get('körper', 30)
                since = (datetime.now() - timedelta(days=days)).strftime('%Y-%m-%d')
                cur.execute(
                    """SELECT date, weight FROM weight_log
                       WHERE profile_id = %s AND date >= %s
                       ORDER BY date DESC""",
                    (profile_id, since)
                )
                variables['weight_data'] = [r2d(r) for r in cur.fetchall()]

            # Nutrition data
            if modules.get('ernährung'):
                days = timeframes.get('ernährung', 30)
                since = (datetime.now() - timedelta(days=days)).strftime('%Y-%m-%d')
                cur.execute(
                    """SELECT date, kcal, protein_g, fat_g, carbs_g
                       FROM nutrition_log
                       WHERE profile_id = %s AND date >= %s
                       ORDER BY date DESC""",
                    (profile_id, since)
                )
                variables['nutrition_data'] = [r2d(r) for r in cur.fetchall()]

            # Activity data
            if modules.get('training'):
                days = timeframes.get('training', 14)
                since = (datetime.now() - timedelta(days=days)).strftime('%Y-%m-%d')
                cur.execute(
                    """SELECT date, activity_type, duration_min, kcal_active, hr_avg
                       FROM activity_log
                       WHERE profile_id = %s AND date >= %s
                       ORDER BY date DESC""",
                    (profile_id, since)
                )
                variables['activity_data'] = [r2d(r) for r in cur.fetchall()]

            # Sleep data
            if modules.get('schlaf'):
                days = timeframes.get('schlaf', 14)
                since = (datetime.now() - timedelta(days=days)).strftime('%Y-%m-%d')
                cur.execute(
                    """SELECT date, sleep_segments, source
                       FROM sleep_log
                       WHERE profile_id = %s AND date >= %s
                       ORDER BY date DESC""",
                    (profile_id, since)
                )
                variables['sleep_data'] = [r2d(r) for r in cur.fetchall()]

            # Vitals data
            if modules.get('vitalwerte'):
                days = timeframes.get('vitalwerte', 7)
                since = (datetime.now() - timedelta(days=days)).strftime('%Y-%m-%d')

                # Baseline vitals
                cur.execute(
                    """SELECT date, resting_hr, hrv, vo2_max, spo2, respiratory_rate
                       FROM vitals_baseline
                       WHERE profile_id = %s AND date >= %s
                       ORDER BY date DESC""",
                    (profile_id, since)
                )
                variables['vitals_baseline'] = [r2d(r) for r in cur.fetchall()]

                # Blood pressure
                cur.execute(
                    """SELECT measured_at, systolic, diastolic, pulse
                       FROM blood_pressure_log
                       WHERE profile_id = %s AND measured_at >= %s
                       ORDER BY measured_at DESC""",
                    (profile_id, since + ' 00:00:00')
                )
                variables['blood_pressure'] = [r2d(r) for r in cur.fetchall()]

            # Mental/Goals (no timeframe, just current state)
            if modules.get('mentales') or modules.get('ziele'):
                # TODO: Add mental state / goals data when implemented
                variables['goals_data'] = []

    # Execute prompt
    return await execute_prompt(prompt_slug, variables, openrouter_call_func, enable_debug)


async def execute_workflow_prompt(
    prompt: Dict,
    variables: Dict[str, Any],
    openrouter_call_func,
    enable_debug: bool = False,
    catalog: Optional[Dict] = None
) -> Dict[str, Any]:
    """
    Execute a workflow-type prompt (graph-based execution).

    Phase 2-4: Sequenzielle Workflow-Execution, conditional branching, path consolidation
    Phase 5: Graph aus ai_prompts.graph_data (nicht workflow_definitions)

    Args:
        prompt: Prompt dict from database (must have 'graph_data' field)
        variables: Dict of variables for placeholder replacement
        openrouter_call_func: Async function(prompt_text, model) -> response_text
        enable_debug: If True, include debug information in response
        catalog: Optional placeholder catalog

    Returns:
        Dict with execution results:
        {
            "type": "workflow",
            "execution_id": "...",
            "status": "completed" | "failed",
            "aggregated_result": {...},
            "node_states": [...],  # Only if enable_debug=True
            "error": "..."         # Only if status=failed
        }
    """
    from workflow_executor import execute_workflow

    # Phase 5: Graph aus ai_prompts.graph_data
    graph_data = prompt.get('graph_data')
    if not graph_data:
        raise HTTPException(400, "Workflow-Prompt fehlt 'graph_data' Feld")

    # Execute workflow (mit graph_data statt workflow_id)
    result = await execute_workflow(
        graph_data=graph_data,  # NEU: Direkt graph_data übergeben
        profile_id=variables.get('profile_id', 'unknown'),  # From context
        variables=variables,
        openrouter_call_func=openrouter_call_func,
        enable_debug=enable_debug
    )

    # Convert ExecutionResult to dict for API response
    response = {
        "type": "workflow",
        "execution_id": result.execution_id,
        "status": result.status,
        "aggregated_result": result.aggregated_result
    }

    if enable_debug:
        response["node_states"] = [s.model_dump() for s in result.node_states]

    if result.error:
        response["error"] = result.error

    return response